Glossary
What Is X-VLA?
Last reviewed: October 2026
Short answer
X-VLA is a published cross-embodiment vision-language-action model from arXiv 2510.10274. Per the abstract, it gives each distinct data source its own learnable embeddings, called soft prompts, in a flow-matching architecture of standard Transformer encoders. Its 0.9B version is evaluated on 6 simulation environments and 3 real-world robots. It is third-party research; Firsthand is not affiliated.
01
What is X-VLA?
X-VLA is introduced in arXiv 2510.10274 (v1, 11 Oct 2025) by Jinliang Zheng, Jianxiong Li and 13 co-authors. The project page lists the Institute for AI Industry Research (AIR) at Tsinghua University, Shanghai Artificial Intelligence Laboratory and Peking University. Read: the arXiv abstract page, the project page and the GitHub README; the full paper text was not read.
02
Where does Firsthand fit next to X-VLA?
Third-party research; Firsthand is not affiliated and claims no use, training or compatibility with it. Firsthand captures first-person human video, not robot data, fast, on demand and to a written spec: depth derived from a calibrated stereo rig, 21 keypoints per hand with a visibility flag, wrist trajectories and action segments, plus one task instruction per episode.
03
What problem does X-VLA address?
The abstract says generalist VLA models rely on training across diverse robotic platforms with large-scale, cross-embodiment, heterogeneous datasets. The project page adds that naively training on heterogeneous data leads to degradation.
04
How do soft prompts work in X-VLA?
Per the abstract, X-VLA brings prompt learning into cross-embodiment robot learning with minimally added parameters, using separate sets of learnable embeddings for each distinct data source. These act as embodiment-specific prompts that help the model use varying cross-embodiment features. The architecture is flow-matching-based and relies exclusively on soft-prompted standard Transformer encoders, which the authors credit with scalability and simplicity.
The project page says the encoding pipeline disentangles high- and low-dimensional input streams, crediting better training stability.
05
What does the X-VLA code expect as input and output?
The README describes a server-client design. A client sends a request with proprioception, a language instruction, a primary RGB image, optional extra camera views, a domain_id for the embodiment and a flow-matching step count. Control uses a unified EE6D end-effector space: a 20-D vector of 3 position values, 6 rotation values, 1 gripper value and 10 padding values, zero-padded for single-arm robots.
06
What does the X-VLA release include?
The README lists a pretrained foundation checkpoint, X-VLA-Pt, checkpoints fine-tuned for embodiments such as Franka, Google Robot, WidowX and Agilex, and LoRA adapters, under the Apache License 2.0. It says X-VLA is integrated into LeRobot. The project page also introduces Soft-Fold, a cloth-folding dataset it calls the first open-source high-quality one, and the authors report using 1,200 episodes for it. The README lists per-checkpoint scores; this page repeats none.
07
How does X-VLA differ from HPT, UniVLA and cross-embodiment learning?
Cross-embodiment learning is the training strategy: one model learns from several embodiments. The site’s HPT entry describes a shared trunk with embodiment-specific tokenizers, and its UniVLA entry describes task-centric latent actions learned from video. X-VLA’s stated move is per-source soft prompts on standard Transformer encoders. The table ranks none of them.
08
What do the X-VLA sources not state?
The full paper’s method detail was not read, so the pretraining data mix and sizes, prompt dimensions and vision-language backbone are Not stated in the text read. The venue differs: the README says ICLR 2026, while the arXiv comments say preprint, technical report. Whether first-person human video would suit X-VLA is not said.
09
What human data pairs with cross-embodiment VLA research like X-VLA, and how do you get it custom-collected?
X-VLA’s stated inputs are robot proprioception, images, a language instruction and a domain ID, and it outputs robot end-effector actions. Firsthand delivers no robot proprioceptive state or robot actions. It records first-person human video to a written spec: depth from a calibrated stereo rig, 21 keypoints per hand with a visibility flag, wrist trajectories, action segments and one task instruction per episode. Confirm tasks and signals in a written spec on a Spec Pilot. Median time to first validated batch: 48 hours where coverage is deep, three to four weeks where thin. Formats include RLDS, LeRobot format and WebDataset; matching X-VLA’s fields is the buyer’s to check.
10
X-VLA setup, as stated in the official sources
| Item | What the official sources state |
|---|---|
| Contribution (abstract) | Embodiment-specific soft prompts: one set of learnable embeddings per data source |
| Architecture (abstract) | Flow-matching-based, standard Transformer encoders only |
| Evaluation (abstract) | 0.9B model on 6 simulation environments and 3 real-world robots |
| Control space (README) | EE6D end-effector, 20-D vector |
| Release (README) | Pretrained and fine-tuned checkpoints, LoRA adapters, LeRobot integration |
| Licence (README) | Apache License 2.0 |
| Pretraining data mix, backbone | Not stated in the text read |
| Firsthand relationship | None |
12
X-VLA beside HPT and UniVLA
| Item | X-VLA (sources read) | HPT (site entry) | UniVLA (site entry) |
|---|---|---|---|
| Stated focus | Soft prompts for heterogeneous cross-embodiment robot data | A shared transformer trunk pre-trained across embodiments | Cross-embodiment policies learned from video |
| Telling embodiments apart | Learnable embeddings per data source; a domain_id in the README client | Embodiment-specific tokenizers for proprioception and vision | Latent actions shared by robot and human video |
| Action output | Flow-matching generation of EE6D actions | Heads map trunk tokens to control | A lightweight head decodes latent actions to robot actions |
| What it is | A model design | A model design | A model design |
13
What X-VLA consumes beside what the Firsthand site states
| Aspect | X-VLA (sources read) | What the Firsthand site states |
|---|---|---|
| Images | A primary RGB image, optional extra views | Head and wrist camera video; depth from a calibrated stereo rig |
| Proprioception | Part of the client payload | Robot proprioceptive state is not delivered |
| Language | A language instruction string | One task instruction per episode |
| Embodiment ID | domain_id integer | Human first-person capture; no robot embodiment |
| Actions | EE6D end-effector actions | Hand pose, wrist trajectories and action segments; no robot actions |
| Format | meta.json plus domain handlers (lerobot, h5py, scattered) | RLDS, LeRobot format, WebDataset; fit to X-VLA handlers Not stated |
14
Frequently asked questions
- What is X-VLA?
- A cross-embodiment vision-language-action model (arXiv 2510.10274) that gives each data source its own learnable soft prompts in a flow-matching, Transformer-encoder architecture.
- How do soft prompts work in X-VLA?
- Per the abstract, separate sets of learnable embeddings for each distinct data source act as embodiment-specific prompts, with minimally added parameters.
- How is X-VLA different from HPT and UniVLA?
- X-VLA uses per-source soft prompts on standard Transformer encoders. The site entries describe HPT as a shared trunk with embodiment-specific tokenizers and UniVLA as learning task-centric latent actions from video.
- Is X-VLA open source?
- The README lists the Apache License 2.0 and releases code, checkpoints and LoRA adapters. The pretraining data mix is Not stated in the text read.
- Does Firsthand use X-VLA, and can I request a custom dataset?
- Firsthand is not affiliated and claims no use, training or compatibility. It delivers human first-person video, not robot proprioception or actions. You can request a custom dataset, confirmed first on a Spec Pilot.
Related terms